Spatiotemporal analysis of plague risk in Tibet: Multi-source data-driven ensemble model development and validation

Abstract

Background

Plague, caused by the highly lethal bacterium Yersinia pestis, remains a critical public health threat. Tibet's unique alpine environment fosters sensitive ecosystems, yet plague dynamics there are understudied. This study reveals environmental drivers and distribution evolution of plague foci in Tibet.

Methods

Using Tibet plague surveillance data (2000−2021), we integrated eight environmental variables (maximum temperature, precipitation, normalized difference vegetation index, land use and land cover, soil moisture, slope, aspect, and population spatial distribution). Ten algorithms via BIOMOD2 modeled plague risk across four phases (2000–2004, 2005–2009, 2010–2015, 2016–2021), validated with 2022–2023 data. We simulated risk distribution, calculated area/centroid changes, and elucidated spatiotemporal evolution.

Results

The ensemble model (EM) showed excellent external validation (2022: AUC = 0.98, TSS = 0.92, Kappa = 0.88; 2023: AUC = 0.99, TSS = 0.99, Kappa = 0.99). Key drivers were population, precipitation, max temperature, and NDVI. The EM revealed stage-wise centroid migration. High-risk areas concentrated in Lhasa, eastern Shigatse, and eastern Nagqu during 2000–2004, contracting spatially 2005–2009 with northeast centroid shift. Post-2010, the centroid shifted southeast then northwest, linked to anthropogenic activities and climate variations.

Conclusion

Multi-model integration and spatiotemporal simulation revealed plague foci distribution and centroid evolution in Tibet, capturing human-climate impacts, providing a scientific basis for targeted prevention in high-risk areas.

Highlights

• •First fine-scale plague risk analysis for Tibet's sensitive alpine ecosystems.
• •BIOMOD2 ensemble model (10 SDMs) achieves AUC ≥ 0.98, offers replicable zoonotic framework.
• •Key drivers POP, PRCP, TMAX, NDVI; plague foci migrate, link humans with climate.

1. Introduction

Plague, caused by the gram-negative bacterium Yersinia pestis (Y. pestis), represents a fulminant natural focal disease with profound historical significance [1]. Having precipitated three global pandemics resulting in hundreds of millions of fatalities, plague natural foci persist across all continents except Oceania and Antarctica [2]. These foci emerge through synergistic interactions between Y. pestis, rodents (e.g., the Marmota himalayana, hereinafter referred to as “marmot”), flea vectors (Callopsylla dolabris and Oropsylla silantiewi), and specific geographical landscapes that sustain plague-enzootic ecosystems [3], [4]. In nature, plague usually begins with an epidemic among animals, which then spills over to humans. However, in some newly identified plague foci, the disease tends to first emerge in human populations, only later being confirmed among animals. Once plague occurs in humans, it typically presents as an acute disease with a high fatality rate and severe symptoms. Therefore, actively preventing and effectively controlling the risk of plague outbreaks is essential for safeguarding public health, and effective monitoring and identification of plague foci are crucial to achieving this goal.

Globally, plague natural foci demonstrate remarkable biogeographic diversity, with China alone hosting 12 distinct focus types [4]. Among these, marmots hibernate from October each year and awakens in April of the following year [5]. Plague outbreaks are closely related to the activity cycles of rodent hosts, usually peaking in June and July [5]. The occurrence, outbreak, and spread of plague are intricately linked to the ecological environment of plague foci. Therefore, analyzing the environment of these foci can effectively predict the risk of plague transmission.

Species distribution models (SDMs) statistically associate species distribution data with environmental variables that limit the species' distribution, establishing a relationship between the species' distribution and environmental factors, which can then be used to predict the geographical distribution or potential habitat of the species [6]. In recent years, due to their strong predictive capability, SDMs have been widely applied by scientists worldwide to analyze changes in species' spatial distribution patterns under climate change scenarios [6], to protect endangered species scientifically [7], and to expand research into the spatial epidemiology of natural zoonotic diseases. These models have been extensively used for the prediction and early warning of diseases such as Ebola virus [8], schistosomiasis [9], and hemorrhagic fever with renal syndrome [10]. In plague control, researchers have also used SDMs to study the prediction of potential risk areas for human plague in relation to meteorological and environmental factors, as well as the suitable habitat areas of major plague host animals. For instance, Holt et al. [11] projected that by 2050, meteorological factors might reduce the risk of plague in southern California, while simultaneously increasing the risk along the northern coast and mountain ranges.

Conventional studies often rely on single-algorithm approaches, yet growing evidence reveals substantial inter-model variability due to differing theoretical assumptions and algorithmic biases [12]. To address this uncertainty, ensemble modeling frameworks like BIOMOD2 [13] integrate 10 advanced modeling techniques, including Surface Range Envelope (SRE), Generalized Linear Models (GLM), Generalized Additive Models (GAM), Classification Tree Analysis (CTA), Artificial Neural Networks (ANN), Generalized Boosted Models (GBM), Random Forest (RF), Flexible Discriminant Analysis (FDA), Multivariate Adaptive Regression Splines (MARS) and Maximum Entropy (MaxEnt) - achieving superior predictive accuracy, particularly for small-sample datasets [14], [15]. This makes plague risk prediction using BIOMOD2 more precise in delineating risk zones.

As the primary component of the Qinghai-Tibet Plateau, Tibet's distinctive high-altitude cryogenic geomorphological patterns and steep vertical climatic gradients have fostered the world's highest-elevation and most ecologically sensitive natural plague foci of the marmots. However, current research predominantly concentrates on pan-Qinghai-Tibetan Plateau investigations, while systematic exploration remains lacking regarding predictive modeling and early warning systems for plague epidemics specific to Tibet, particularly the mechanisms of plague dynamics in response to climatic changes.

This study focuses on the plague foci in the Tibet Autonomous Region. Based on monitoring data from 2000 to 2021, we use spatiotemporal dynamic modeling (with four periods: 2000–2004, 2005–2009, 2010–2015, and 2016–2021) to explore the evolution of the distribution patterns of these foci. Furthermore, two independent datasets between 2022 and 2023 are used to validate the model's extrapolation capability. Specifically, our objectives are to: (1) reveal the environmental driving mechanisms of the spatiotemporal distribution of plague foci in the Tibet Autonomous Region; (2) establish a high-precision plague risk prediction model; and (3) assess the potential risk of focal expansion under climate change scenarios, providing spatial decision support for the precise prevention and control of plague in Tibet.

2. Materials and methods

2.1. Study region and occurrence data

2.1.1. Study region

The study area covers the entire Tibet Autonomous Region (26°50′–36°53′N, 78°25′–99°06′E) with a total area of approximately 1,228,400 km², representing the core distribution zone of marmots and a major concentration of natural plague foci. Dominant vegetation types include alpine meadows, alpine steppes, desert steppes, valley shrubs, and coniferous forests, whose distribution patterns are strongly regulated by altitude and precipitation gradients. The region exhibits pronounced heterogeneity in climatic and environmental factors: the southeastern canyon areas receive abundant precipitation with distinct vertical climate zonation, while the northwestern plateau hinterland displays typical alpine-arid characteristics. This geographical diversity provides complex habitat conditions for the formation of host-vector-pathogen ecological system of plague [16], [17].

2.1.2. Occurrence data

All plague surveillance data were obtained from the Plague Prevention and Control Management Information System of the Tibet Autonomous Region Center for Disease Control and Prevention, encompassing animal and human plague monitoring records from 2000 to 2023. To comprehensively understand plague risk drivers, we implemented a two-tiered analytical approach. First, we constructed a model using all occurrence data from 2000 to 2021 to identify overarching, stable environmental influencing factors of plague distribution in Tibet. Second, to capture spatiotemporal dynamics central to our study objectives, we employed a temporal stratification strategy. The 2000–2021 data were divided into four periods (2000–2004: 44 foci, 2005–2009: 58 foci, 2010–2015: 70 foci, 2016–2021: 48 foci), with each period modeling the distribution pattern of plague foci. This segmentation was determined based on: (1) the fact that 2005 marked the release of China's first national plague surveillance protocol [18], which significantly standardized and intensified monitoring efforts across the country, including Tibet; (2) the need to ensure a sufficient number of presence points per period for reliable model training; and (3) the goal of creating periods of approximately equal length to facilitate a clear and balanced comparative analysis of distribution dynamics across stages. To further validate the external predictive ability of the model, we used monitoring and environmental data from 2022 (12 foci) and 2023 (8 foci) to test the model calibrated on the most recent period (2016–2021), as this model best represents the current state of the plague foci system and is therefore most relevant for forward–looking prediction and public health intervention. All plague foci were bacteriologically confirmed and georeferenced via GPS or geocoding. The response variable (Y) for model training was defined as the presence of a natural plague focus. Human plague cases were overlaid onto the predicted risk maps, providing a visual and qualitative assessment of performance. To ensure spatial independence, duplicate records within the same 1 km × 1 km grid cell (environmental data resolution) were removed using QGIS 3.34.1 (spatial accuracy ≤0.1°).

2.2. Selection and processing of environmental variables

2.2.1. Data collection

The distribution of marmots is primarily influenced by climate, vegetation, and topography [19], [20], [21]. We selected 12 environmental variables considered to be associated with the presence of plague natural foci (Table 1). Meteorological data, including average temperature [22], maximum temperature [22], minimum temperature [22], and precipitation [22], were sourced from the National Tibetan Plateau Science Data Center [23], [24], [25], [26]. Normalized Difference Vegetation Index (NDVI) data were obtained from the MOD13A3 dataset regularly released by NASA (Available from: https://www.earthdata.nasa.gov/). Land cover data were derived from the 30 m annual land cover grid data for China (1990–2023) published by Yang Jie et al. [27], which include cropland, forest, grassland, water, barren land, and snow ice. Soil moisture data were obtained from the monthly gap–filled CCI Soil Moisture dataset for China (1982–2020) generated by Sun Hao et al. [28]. This dataset is based on the European Space Agency's (ESA) raw soil moisture product, and gaps were filled using the XGBoost algorithm, which integrates precipitation, surface temperature, NDVI, and other covariates for spatial interpolation of missing values [28]. The dataset was validated with 192 ground stations, with a correlation coefficient of 0.554 and a root mean square error (RMSE) of 0.0977 cm³/cm³, which outperforms the original CCI data. Elevation data were obtained from the Copernicus 30–meter resolution DEM dataset [29], and three derived topographical variables-slope, aspect, and terrain ruggedness index-were calculated using the spatial analysis module of QGIS 3.34.1. Population spatial distribution data were sourced from the LandScan Global Population Database (Available from: https://landscan.ornl.gov/).

Table 1.

Environmental variables used to predict the plague distribution in Tibet.

Abbreviations	Environment variable	Spatial Resolution	Time granularity
TAVG	Average Temperature	1 km	Monthly
TMAX	Maximum Temperature	1 km	Monthly
TMIN	Minimum Temperature	1 km	Monthly
PRCP	Precipitation	1 km	Monthly
NDVI	Normalized Difference Vegetation Index	1 km	Monthly
LULC	Land Use and Land Cover	30 m	Annual
SM	Soil Moisture	0.25°	Monthly
DEM	Digital Elevation Model	30 m	Static
SLOPE	Slope	30 m	Static
ASPECT	Aspect	30 m	Static
TRI	Terrain Ruggedness Index	30 m	Static
POP	Population Spatial Distribution	1 km	Annual

2.2.2. Processing of environmental variables

All environmental variables underwent the following standardization process: (1) Spatial Dimension: All data were resampled to a 1 km resolution using bilinear interpolation (for continuous variables) or nearest neighbor methods (for categorical variables). For modeling purposes, all environmental layers were masked to fit the boundary of the Tibet Autonomous Region. The projection coordinate system was defined as WGS1984. (2) Temporal Dimension: Monthly meteorological and NDVI data were averaged for the marmot activity period (April to October). For annual data, such as population spatial distribution (POP), we directly used the annual mean for each study period. For the categorical land use and land cover (LULC) variable, which exhibits relatively low inter-annual fluctuation, the data from the median year of each period (e.g., 2002 for 2000–2004) were employed. This approach provides a temporally balanced representation of land cover for the entire period, minimizing potential bias that could arise from selecting a year at either temporal extreme when correlating with plague occurrence data. To avoid multicollinearity between variables, which could lead to model overfitting, we calculated the Spearman correlation coefficient for each pair of variables (Fig. 1). From highly correlated pairs (|r| > 0.8), we retained a single variable based on greater biological relevance and to avoid informational redundancy. Specifically, maximum temperature (TMAX) was selected over other temperature metrics based on the premise that extreme heat can be a critical limiting factor for flea vector survival and activity [30], while slope was retained over the terrain ruggedness index (TRI) owing to its clearer ecological interpretability [31]. The final retained variables included: maximum temperature (TMAX), precipitation (PRCP), normalized difference vegetation index (NDVI), land use and land cover (LULC), soil moisture (SM), slope, aspect, and population spatial distribution (POP).

Fig. 1.

The correlation analysis of environmental variable.

2.3. Construction and evaluation of species distribution model

Based on the plague monitoring data from Tibet (2000–2021), we first fitted the data using 10 individual methods available in BIOMOD2. Before constructing the model, it was necessary to process the species distribution data. Biomod2 provides several methods to generate pseudo-absence points from background study data [13]. Using the “random” command, we generated twice as many pseudo–absence data points as presence points for model simulation. This 1:2 ratio is a common and effective practice in species distribution modeling [32], as it provides sufficient data for model training while minimizing the dilution of ecological signal from the presence data. The model training adopted the “user.defined” strategy: specifically, for the Random Forest (RF) and XGBoost models, key hyperparameters were configured based on preliminary tests and common academic practices to optimize model performance and mitigate overfitting; whereas all other modeling algorithms adhered to the default settings of the BIOMOD2 package. The final parameter sets for these models are provided in Supplementary Table 1. For model training and internal validation, we randomly allocated 80% of the sample data to the training set and the remaining 20% to the validation set [33], and this process was repeated five times to avoid the stochastic nature of single model fitting.

We evaluated each individual model using three metrics: Area Under the Curve (AUC), True Skill Statistic (TSS), and Kappa coefficient [34]. AUC represents the area under the Receiver Operating Characteristic (ROC) curve, with values ranging from 0.5 to 1. A value greater than 0.7 indicates a moderate prediction, greater than 0.8 is a good result, and greater than 0.9 represents an excellent prediction. TSS combines sensitivity and specificity, accounting for omission and commission errors. TSS value ranges from −1 to 1, with values between 0.4 and 0.6 indicating moderate fit, between 0.6 and 0.8 indicating good fit, and between 0.8 and 1 indicating excellent fit. The Kappa coefficient incorporates the species distribution probability, sensitivity, and specificity, providing a more robust and conservative accuracy measure by excluding random factors influencing consistency between two variables [35]. Kappa values range from 0 to 1, with higher values indicating better model performance.

We used a weighted averaging method to integrate individual models that met the selected accuracy criteria into an ensemble model [36]. The weight of each model in the ensemble was determined based on its TSS value, and only models with TSS > 0.8 were included for constructing the ensemble model. Relevant code is provided in Supplementary Text 1.

To identify the key environmental drivers of plague risk from the ensemble model, the relative contribution of each environmental variable to the ensemble model was quantified using permutation importance, a robust metric. This metric evaluates the impact of each variable by randomly permuting its values over 10 iterations and measuring the resulting decrease in model performance (based on TSS). A greater decrease indicates a more important variable.

2.4. Geographical analysis

The risk distribution map generated by the ensemble model ranges from 0 to 1000, indicating the probability of plague occurrence. Based on the IPCC assessment report and the classification standards in existing research [37], [38], we divided the risk distribution map into four levels: negligible risk area (0−220), low risk area (220–500), moderate risk area (500–750), and high risk area (750–1000). We then used the “metric.binary” method to select the optimal threshold value, constructing binary distribution maps for presence and absence. These resultant binary maps were used for two key geographical analyses: (1) to calculate and compare the areal extent of plague risk zones across the four different periods, quantifying the expansion or contraction of risk zones; and (2) to calculate the centroid of the plague ‘presence’ for each period. The centroid shifts between consecutive periods were then quantified by calculating the Euclidean distance between them in the WGS 1984 coordinate reference system. All analyses in this study were conducted using QGIS 3.34.1 and R 4.4.1.

3. Results

3.1. Epidemiological characteristics of plague in Tibet

The analysis is built upon plague surveillance data spanning from 2000 to 2023. During this period, a total of 25 human plague cases were reported in the Tibet Autonomous Region, resulting in 16 deaths. A total of 666 strains of Yersinia pestis were isolated.

Spatially, the distribution of plague was highly heterogeneous. A total of 54 counties (accounting for 72.97% of all counties in Tibet) were confirmed as natural plague foci based on the isolation of Yersinia pestis. Lhasa, Shigatse, and Shannan City were the predominant endemic areas, collectively accounting for 477 bacterial isolates (71.62% of the total) and 16 human cases (62% of the total).

Plague occurrence demonstrated marked seasonality. Animal plague outbreaks were concentrated between June and August, peaking in July. Human cases occurred primarily from July to September, slightly lagging behind the peak in animal plague, which is consistent with the typical transmission cycle from animal hosts to humans.

The affected individuals were predominantly male (20 cases, 80%), with a mean age of 35.5 years, and herders (20 cases, 80%). The main reported transmission routes included contact with infected patients (12 cases, 48%) and contact with or handling of marmots (6 cases, 24%). Pneumonic plague was the predominant clinical form (13 cases, 52%), followed by bubonic plague (4 cases, 16%).

3.2. Model performance and predictive accuracy

The ensemble model for the entire 2000–2021 period demonstrated high predictive accuracy (AUC = 0.95, TSS = 0.88, KAPPA = 0.82), capturing the overarching distribution pattern of plague risk in Tibet. The time-staged ensemble models (EMs) also showed consistently high performance in internal 5-fold cross-validation across all four periods (Supplementary Table 2), confirming the robustness of our stage-specific analysis.

Furthermore, the optimal model constructed from the 2016–2021 training period was applied to independent environmental datasets from 2022 and 2023 for dual validation. The validation results showed that the ensemble model (EM) demonstrated exceptional performance across both validation phases: AUC = 0.98, TSS = 0.92, and Kappa = 0.88 for 2022; AUC = 0.99, TSS = 0.99, and Kappa = 0.99 for 2023 (Table 2), indicating strong temporal generalizability. Among individual models, RF achieved the best performance in 2022 (AUC = 0.97, TSS = 0.88, Kappa = 0.85), while GBM performed optimally in 2023 (AUC = 0.99, TSS = 0.95, Kappa = 0.94). SRE model showed no predictive capability in either phase (AUC = 0.5).

Table 2.

Comparative performance evaluation of SDMs (2022 and 2023).

Model	2022			2023
	AUC	TSS	KAPPA	AUC	TSS	KAPPA
EM	0.98	0.92	0.88	0.99	0.99	0.99
RF	0.97	0.88	0.85	0.94	0.87	0.82
GBM	0.93	0.76	0.78	0.99	0.95	0.94
GAM	0.91	0.77	0.78	0.98	0.96	0.94
FDA	0.90	0.79	0.80	0.82	0.64	0.56
GLM	0.89	0.71	0.71	0.93	0.82	0.77
XGBOOST	0.87	0.67	0.70	0.97	0.85	0.87
CTA	0.83	0.67	0.73	0.93	0.88	0.86
MAXENT	0.81	0.62	0.67	0.93	0.85	0.82
ANN	0.72	0.45	0.46	0.93	0.85	0.80
SRE	0.5	0	0	0.5	0	0

3.3. Spatial distribution and changes of natural plague foci

Building upon the robust model performance, we uncovered significant spatiotemporal dynamics in plague risk. The spatial distribution of plague foci in Tibet during different periods was visualized using QGIS, with the projection coordinate system defined as WGS 1984. Fig. 2 shows the simulated distribution of plague foci in Tibet, where the color gradient from green to red represents the increasing risk, and blue points indicate human plague cases during that period. From Fig. 2a, it can be seen that during 2000–2004, the high-risk areas (risk probability >750) covered approximately 42,136 km² (3.4% of the region) and showed a clustered pattern, primarily concentrated in the eastern part of Lhasa, the eastern part of Shigatse, the northwest of Shannan, and the eastern part of Nagqu. By the period of 2005–2009 (Fig. 2b), the high-risk areas expanded to 47,584 km² (3.9%), with new risk clusters emerging in the Qamdo region. In the subsequent periods, the high-risk zone contracted substantially, declining to 23,260 km² (1.9%) in 2010–2015 (Fig. 2c) and further to 20,353 km² (1.7%) in 2016–2021 (Fig. 2d). The complete areal statistics for all risk levels across the four periods are detailed in Supplementary Table 3.

Fig. 2.

Risk maps of natural plague foci in Tibet over different periods.

These spatial patterns are consistent with the spatial overlap analysis conducted using binary distribution maps from the four periods (Supplementary Fig. 1). Quantitative analysis of these binary maps further revealed distinct dynamics in the plague risk areas (Supplementary Table 4). The most notable expansion occurred between 2000 and 2004 and 2005–2009, with 87,167 km² (7.1%) of new areas becoming at-risk. In contrast, the period from 2005 to 2009 to 2010–2015 was characterized by substantial contraction, with 87,722 km² (7.1%) of previously at-risk areas transitioning to negligible risk.

Importantly, these model-predicted high-risk areas showed clear spatial coupling with the distribution of confirmed human plague cases during the same period. Furthermore, the external prediction results derived from the model built during the 2016–2021 training period (Supplementary Fig. 2) were in good agreement with the actual distribution areas of plague foci.

3.4. Center shift of plague natural foci

Based on QGIS 3.34.1, an analysis of the centroids of plague risk distributions during different periods showed a clear trend of centroid migration (Fig. 3). From the period of 2000–2004 to 2005–2009 (Stage 1–2), the centroid migrated northeast by 79.39 km. This was followed by a pronounced southeastward shift of 163.85 km during 2005–2009 to 2010–2015 (Stage 2–3). Finally, the centroid reversed direction and moved northwestward by 103.03 km in the transition to 2016–2021 (Stage 3–4). Overall, centroid fluctuations concentrated in central Tibet.

Fig. 3.

Centroid migration of the risk distribution of natural plague foci in Tibet across different stages.

3.5. Dynamic drivers underlying the spatiotemporal patterns

3.5.1. Predictor variable contributions

The relative contribution of environmental drivers was quantified and compared between the entire study period (2000–2021) and the four sub-periods using the ensemble model (Table 3). POP, PRCP, TMAX, and NDVI consistently exerted substantial influences. Stage-specific variable importance revealed that while POP, PRCP, and NDVI were consistently influential, the dominance of maximum temperature (TMAX) in the second stage (2005–2009, 46.98%) was particularly notable. LULC, slope, and aspect exhibited minimal impacts on plague focus distribution.

Table 3.

Ranking of the contribution of variables to plague risk distribution across the entire period and the four sub-periods.

Variable	Entire period (2000–2021)	Stage 1 (2000–2004)	Stage 2 (2005–2009)	Stage 3 (2010–2015)	Stage 4 (2016–2021)
POP	1 (55.18% ± 2.09%)	1 (44.81% ± 3.62%)	3 (16.21% ± 1.35%)	1 (73.38% ± 5.44%)	1 (52.73% ± 4.71%)
PRCP	2 (17.24% ± 1.41%)	2 (36.02% ± 3.20%)	2 (18.15% ± 1.54%)	3 (9.10% ± 1.27%)	3 (16.56% ± 2.17%)
NDVI	3 (12.53% ± 0.70%)	3 (9.29% ± 1.06%)	4 (14.40% ± 1.03%)	2 (10.02% ± 1.84%)	4 (4.18% ± 0.75%)
TMAX	4 (11.19% ± 0.66%)	5 (3.12% ± 0.39%)	1 (46.98% ± 3.83%)	4 (5.49% ± 0.58%)	2 (17.49% ± 2.69%)
SM	5 (1.55% ± 0.12%)	4 (3.36% ± 0.39%)	7 (0.72% ± 0.06%)	6 (0.50% ± 0.14%)	6 (3.66% ± 0.61%)
LULC	6 (1.18% ± 0.13%)	7 (1.19% ± 0.17%)	5 (2.17% ± 0.21%)	7 (0.49% ± 0.04%)	7 (0.93% ± 0.12%)
SLOPE	7 (0.66% ± 0.06%)	6 (1.23% ± 0.16%)	8 (0.59% ± 0.05%)	5 (0.74% ± 0.31%)	5 (4.14% ± 0.71%)
ASPECT	8 (0.46% ± 0.03%)	8 (0.99% ± 0.04%)	6 (0.78% ± 0.07%)	8 (0.27% ± 0.03%)	8 (0.32% ± 0.09%)

3.5.2. Interpretation of environmental drivers

The direction and form of the key environmental drivers were elucidated by analyzing GLM coefficients and ensemble model response curves (Supplementary Table 5; Supplementary Fig. 3). POP demonstrated a robust and consistent positive association with plague risk across all periods (P < 0.05), with significant quadratic terms in multiple periods indicating consistent nonlinearity. PRCP exhibited temporal heterogeneity, showing significant positive linear effects during Stage 1(2000–2004) and Stage 3(2010–2015), with negative quadratic terms in these two stages revealing nonlinear, unimodal relationships.

NDVI displayed significant unimodal relationships during the entire period and Stage 2, characterized by positive linear and negative quadratic terms. In Stage 3, NDVI showed a significant nonlinear relationship with a positive quadratic term. TMAX exhibited a temporally evolving effect, with Stage 4 showing a significant nonlinear, single-peaked relationship (positive linear and negative quadratic terms, P = 0.03 and P = 0.05), with optimum risk at approximately 20 °C.

4. Discussion

This study employed the BIOMOD2 ensemble modeling framework to analyze plague risk distribution in Tibet by integrating multiple environmental variables. The multi-algorithm fusion strategy significantly improved predictive accuracy, primarily through inter-model calibration that reduced instability and overfitting [39]. Importantly, combining algorithms with divergent assumptions—such as the non-linear capture of RF versus the smoothing functions of GAM—allowed the ensemble model to reveal both abrupt ecological thresholds and continuous environmental constraints. This approach provides insights into complex, non-linear host–environment interactions that exceed the interpretive capacity of any single model.

The high predictive accuracy of our ensemble model is substantiated by its close alignment with key epidemiological features of plague in Tibet observed between 2000 and 2023. During this period, the spatial clustering of bacterial isolates in Lhasa, Shigatse, and Shannan corresponded closely to the high-risk areas identified by the model, providing empirical validation of the predicted risk surface. The predominance of cases among young male herders suggests behaviorally mediated exposure pathways, which likely contribute to the positive association between POP and plague risk detected by the model.

A spatial mismatch was observed between some known plague foci and model-predicted high-risk areas. This discrepancy likely reflects two factors. The model captures regions of long-term environmental suitability, while the actual establishment of foci is often stochastic, influenced by short-term climatic anomalies or local conditions not captured in multi-year averages. Historical surveillance may also be geographically biased, limiting the known distribution of foci. As a result, high-risk areas without recorded foci may indicate surveillance gaps or undetected enzootic transmission, whereas foci in lower-risk zones may result from localized detection efforts. These patterns demonstrate the practical value of ensemble modeling: it not only validates known distributions but also identifies regions where ecological suitability diverges from surveillance records, informing proactive surveillance strategies.

Building on this validated predictive framework, our two-tiered analytical approach provided complementary insights into plague ecology in Tibet. While the whole-period model identified POP and PRCP as stable drivers, stage-specific models revealed pronounced temporal dynamics. These dynamics demonstrate a complex interplay of climatic and anthropogenic factors, with POP, PRCP, TMAX, and NDVI showing distinct functional responses that help explain the observed shifts in plague distribution.

The migration trajectory of the plague risk centroid reflects the dynamic interplay of multiple environmental drivers. The northeastward shift (2000–2004 to 2005–2009) coincided with TMAX emerging as the dominant factor (46.98%), suggesting that climatic warming provided more favorable conditions for marmots and flea vectors [40]. The ensemble model identified a unimodal response of plague risk to TMAX, peaking at 20 °C. This optimal temperature aligns with experimental studies showing peak flea transmission efficiency near 23 °C and inhibition above 30 °C [41], appearing repeatedly across temporal stages. This pattern indicates a robust thermal optimum for vector competence and illustrates how multi-model integration strengthens ecological inference.

The subsequent southeastward shift of the centroid from 2005 to 2009 to 2010–2015 reflects a multi-driver regime, with climatic and anthropogenic factors acting synergistically. POP remained the dominant variable, though its effect was likely modulated by higher precipitation and more abundant vegetation (NDVI) in southeastern Tibet, creating favorable microenvironments for flea reproduction and persistence [42]. At the same time, human activities in these regions may have further amplified local transmission risk by providing supplemental food resources, such as grain stores and livestock fodder, which can elevate rodent carrying capacity and promote population growth [43]. These findings align with a study in Tanzania showing that human activity spaces were significantly correlated with flea distributions, particularly in zones of frequent agricultural and foraging activities [44].

Precipitation in Tibet generally showed a negative or unimodal relationship with plague risk, contrasting with Inner Mongolia, where summer rainfall promotes risk [45]. This difference reflects how identical drivers operate through distinct ecological mechanisms across ecosystems. In Tibet's alpine environment, the ensemble model captured a unimodal or inhibitory effect, suggesting that excessive rainfall may waterlog marmot burrows and create unfavorable flea microclimates, unlike the vegetation-mediated enhancement seen in arid grasslands. Vegetation (NDVI) also showed a context-dependent, non-linear response, peaking around 0.6—higher than reported for the great gerbil in the Junggar Basin [46]. By integrating multiple algorithms, the model identified these thresholds and clarified how moderate vegetation supports marmot habitat, while overly dense growth may limit plague transmission.

The northwestward migration in the most recent period (2016–2021) highlights the dual role of human influence. Regions with higher human density receive intensified surveillance, increasing the likelihood of plague detection, and human settlements can elevate local transmission risk by providing additional food resources that boost host carrying capacity. Human activities can also propagate plague foci into new areas. Intensive land use, such as grassland reclamation and livestock grazing, degrades and fragments marmot habitats [47], while rapid tourism development around Lhasa and other areas may displace marmots toward less disturbed northwestern grasslands [48]. Tree-based algorithms (RF, GBM) captured threshold-like responses of plague risk to population density, reflecting non-linear effects of habitat fragmentation and host displacement. In contrast, generalized linear models (GLM) showed a continuous, proximity-driven increase. Together, these complementary algorithmic insights indicate that population density acts not merely as a linear predictor but as a proxy for intertwined socio-ecological processes—both intensifying local transmission and driving geographic redistribution of plague foci.

From a public health perspective, these shifting geographic hotspots emphasize the need for adaptive surveillance strategies that integrate real-time climate data and human activity mapping in Tibet. By integrating real-time remote sensing data, interventions can be more precisely targeted at high-risk areas. It is particularly important to implement a gradient-based control strategy in areas such as population-dense zones, precipitation transition zones, temperature threshold zones, and areas with anomalous NDVI values. For example, controlling host density in high NDVI areas and conducting health education in hotspots of human migration can form an “environment-host-human” triad for a more effective control network.

The strong influences of population density, seasonal precipitation, maximum temperature, and NDVI observed in Tibet may provide a conceptual reference for other marmot-associated plague foci. In adjacent Qinghai Province, where the same host species (Marmota himalayana) occurs, similar environmental drivers are likely relevant, though their relative contributions may vary locally. Beyond the Qinghai–Tibet Plateau, these findings may also inform studies of other high-altitude grasslands maintained by different marmot species. For example, research in the Tien Shan Mountains of Kyrgyzstan identified the grey hamster (Marmota baibacina) as a key plague host in a comparable mountainous steppe [49]. While specific climate thresholds or human–animal interaction patterns may differ across regions, linking plague risk to climatic, vegetative, and anthropogenic factors provides a broadly applicable framework. In markedly different ecological regions, environmental associations are expected to differ substantially, emphasizing the need for localized models based on region-specific surveillance data.

Despite the high predictive performance of the ensemble model, there are still some limitations. First, this study did not include certain risk factors that cannot be mapped, such as natural barriers that may impede the spread of marmots, flea vectors, or Yersinia pestis. Second, although marmot density represents a biologically crucial driver of plague dynamics, comprehensive monitoring data were unavailable across all endemic counties and years, precluding its direct inclusion in our models. Third, our approach modeled synchronous climate-plague relationships without explicitly testing potential time-lag effects, which may have been attenuated by multi-year averaging and the pronounced influence of same-season climatic conditions. Future research should therefore prioritize systematic host and vector monitoring, integrate higher temporal-resolution data to examine lagged responses, and establish dynamic data acquisition protocols to strengthen long-term ecological inference.

5. Conclusion

This study employs an ensemble modeling framework and a time-stratified analytical approach to unravel the spatiotemporal dynamics of plague foci. It effectively captures the multidimensional impacts of environmental variables, including human activities and climate change, on the distribution of plague risks. By adapting to evolving environmental conditions, this framework accurately identifies risk variation trends, thereby optimizing resource allocation and enhancing early warning systems for prevention and control. These findings provide valuable insights for improving plague management strategies in Tibet.

CRediT authorship contribution statement

Luo Guo: Writing – review & editing, Writing – original draft, Investigation, Formal analysis, Data curation, Conceptualization. Xiaoyan Zhang: Writing – review & editing, Writing – original draft, Supervision, Methodology, Investigation, Conceptualization. Zhan Lin: Writing – original draft, Investigation, Data curation, Conceptualization. Zhuang Cui: Investigation, Data curation. Changping Li: Investigation, Data curation. Su Wu: Writing – review & editing, Project administration, Methodology, Formal analysis, Conceptualization.

Funding

This study was supported by the National Science National Natural Science Foundation of China (No. 72374153), Key Science and Technology Research and Development Plan Project of Xizang Autonomous Region, China (XZ202502ZY0024).

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

Supplementary material

Data availability

The data underlying this article cannot be shared publicly due to the sensitive nature of the geographic location information. The data will be shared on reasonable request to the corresponding author.